[Colloquium] REMINDER: 11/18 TTIC Colloquium: David Forsyth, University of Illinois at Urbana-Champaign

[Mary Marre]

*When:*      Monday, November 18th at 11:00 am



*Where:*     TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 526



*Who: *       David Forsyth, University of Illinois at Urbana-Champaign




*Title:        *Extending Intrinsic Images with Authorable Image
Decompositions

*Abstract: *An intrinsic image decomposition breaks images into components
that explain the image.  Intrinsic components are object properties like
albedo; extrinsic components
are properties of viewing conditions, like shading.  The original intrinsic
image decomposition algorithm, Retinex, decomposes an image into albedo and
shading.  However, there
are many other intrinsic and extrinsic effects one could investigate.  For
example, surface relief produces complex shading effects that are
essentially intrinsic.  As another example,
gloss effects are extrinsic, but very different from shading effects.
Retinex is “learned” but uses essentially no training data; instead, it
uses a model of the spatial structure of
intrinsic image maps.  Modern practice has been to learn albedo-shading
decompositions from rendered data using a regression strategy.  It is very
difficult to extend this approach
to handle a wider range of intrinsic and extrinsic effects.  I will show
that decompositions into two intrinsic and two extrinsic layers are easily
learned using paradigms — fake
data items that capture important spatial statistics.  I will demonstrate
these decompositions on a wide range of applications and datasets.  Why
this approach works remains somewhat
mysterious, but suggests interesting lines of thought about what neural
networks do.

*Host:* Avrim Blum <avrim at ttic.edu>


For more information on the colloquium series or to subscribe to the
mailing list, please see http://www.ttic.edu/colloquium.php


Mary C. Marre
Administrative Assistant
*Toyota Technological Institute*
*6045 S. Kenwood Avenue*
*Room 517*
*Chicago, IL  60637*
*p:(773) 834-1757*
*f: (773) 357-6970*
*mmarre at ttic.edu <mmarre at ttic.edu>*


On Sun, Nov 17, 2019 at 6:35 PM Mary Marre <mmarre at ttic.edu> wrote:

> *When:*      Monday, November 18th at 11:00 am
>
>
>
> *Where:*     TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 526
>
>
>
> *Who: *       David Forsyth, University of Illinois at Urbana-Champaign
>
>
>
>
> *Title:        *Extending Intrinsic Images with Authorable Image
> Decompositions
>
> *Abstract: *An intrinsic image decomposition breaks images into
> components that explain the image.  Intrinsic components are object
> properties like albedo; extrinsic components
> are properties of viewing conditions, like shading.  The original
> intrinsic image decomposition algorithm, Retinex, decomposes an image into
> albedo and shading.  However, there
> are many other intrinsic and extrinsic effects one could investigate.  For
> example, surface relief produces complex shading effects that are
> essentially intrinsic.  As another example,
> gloss effects are extrinsic, but very different from shading effects.
> Retinex is “learned” but uses essentially no training data; instead, it
> uses a model of the spatial structure of
> intrinsic image maps.  Modern practice has been to learn albedo-shading
> decompositions from rendered data using a regression strategy.  It is very
> difficult to extend this approach
> to handle a wider range of intrinsic and extrinsic effects.  I will show
> that decompositions into two intrinsic and two extrinsic layers are easily
> learned using paradigms — fake
> data items that capture important spatial statistics.  I will demonstrate
> these decompositions on a wide range of applications and datasets.  Why
> this approach works remains somewhat
> mysterious, but suggests interesting lines of thought about what neural
> networks do.
>
> *Host:* Avrim Blum <avrim at ttic.edu>
>
>
> For more information on the colloquium series or to subscribe to the
> mailing list, please see http://www.ttic.edu/colloquium.php
>
>
> Mary C. Marre
> Administrative Assistant
> *Toyota Technological Institute*
> *6045 S. Kenwood Avenue*
> *Room 517*
> *Chicago, IL  60637*
> *p:(773) 834-1757*
> *f: (773) 357-6970*
> *mmarre at ttic.edu <mmarre at ttic.edu>*
>
>
> On Mon, Nov 11, 2019 at 6:35 PM Mary Marre <mmarre at ttic.edu> wrote:
>
>> *When:*      Monday, November 18th at 11:00 am
>>
>>
>>
>> *Where:*     TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 526
>>
>>
>>
>> *Who: *       David Forsyth, University of Illinois at Urbana-Champaign
>>
>>
>>
>>
>> *Title:        *Extending Intrinsic Images with Authorable Image
>> Decompositions
>>
>> *Abstract: *An intrinisc image decomposition breaks images into
>> components that explain the image.  Intrinsic components are object
>> properties like albedo; extrinsic components
>> are properties of viewing conditions, like shading.  The original
>> intrinsic image decomposition algorithm, Retinex, decomposes an image into
>> albedo and shading.  However, there
>> are many other intrinsic and extrinsic effects one could investigate.
>> For example, surface relief produces complex shading effects that are
>> essentially intrinsic.  As another example,
>> gloss effects are extrinsic, but very different from shading effects.
>> Retinex is “learned” but uses essentially no training data; instead, it
>> uses a model of the spatial structure of
>> intrinsic image maps.  Modern practice has been to learn albedo-shading
>> decompositions from rendered data using a regression strategy.  It is very
>> difficult to extend this approach
>> to handle a wider range of intrinsic and extrinsic effects.  I will show
>> that decompositions into two intrinsic and two extrinsic layers are easily
>> learned using paradigms — fake
>> data items that capture important spatial statistics.  I will demonstrate
>> these decompositions on a wide range of applications and datasets.  Why
>> this approach works remains somewhat
>> mysterious, but suggests interesting lines of thought about what neural
>> networks do.
>>
>> *Host:* Avrim Blum <avrim at ttic.edu>
>>
>>
>> For more information on the colloquium series or to subscribe to the
>> mailing list, please see http://www.ttic.edu/colloquium.php
>>
>>
>> Mary C. Marre
>> Administrative Assistant
>> *Toyota Technological Institute*
>> *6045 S. Kenwood Avenue*
>> *Room 517*
>> *Chicago, IL  60637*
>> *p:(773) 834-1757*
>> *f: (773) 357-6970*
>> *mmarre at ttic.edu <mmarre at ttic.edu>*
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20191118/b1e60739/attachment-0001.html>